Rival good

My use reduces the amount of the good for others

Non-rival good

My use doesn’t reduce the amount of the good for others

Excludable good

People can be prevented from using or accessing the good

Non-excludable good

People cannot be prevented from using or accessing the good

Public policy

Government action or inaction in response to public problems

Public problems

Conditions the public sees as unacceptable and in need of intervention

Why do values matter?

Disagreements between which problems matter, what policies should address them and how the policies should be implemented

Values

Equity, efficiency, transparency, liberty, security, accountability and responsiveness

What values compete with each other?

Equity and efficiency Liberty and security Accountability and responsiveness Transparency and security

AI

Computer systems designed to perform tasks that typically require human intelligence

Machine learning

Type of AI where computers learn patterns from data instead of being programmed with rules

Supervised learning

Learns from labeled examples, adjusts patterns and tests on held-out data to evaluate predictions

Unsupervised learning

Finds patterns in unlabeled data, groups or organizes the data to interpret the results

Deep learning

A type of machine learning that uses large models and neural networks

Decision trees

Used for categorical data, splits data into groups based on features and uses a series of questions to predict

Regression

Used for continuous data, uses mathematical relationships between features to find a relationship that best fits the data

Traditional machine learning

We create the feature vectors, choose the features that matter most and the computer uses those to classify things

Neural network

We build the network structures and provide examples, but the system figures out what is what

Neural network learning process

Measure the error → adjust the model → repeat → improve

Node

A neuron that receives information from the previous layer, processes it and sends an output to the next layer

Weight

How much each input matters

Bias

Adds an adjustment to the neuron before it produces the input

Activation function

Determines the neuron’s output based on the information it receives

Forward pass

Run the training data through the model to make a prediction

Loss function

Measures how wrong the prediction was

Backpropagation

Sends the error backward through the network so the model can adjust the weights and biases

Gradient descent

Determines how much to adjust the weights and biases

How adjustments happen

Guess → error → adjust → repeat

Generative AI

Learns patterns in the data to create new examples that fit those patterns

General Adversarial Networks (GANs)

Uses neural networks (generator and discriminator), creates new images from random inputs

Diffusion models

Uses one neural network, generates images by gradually removing noise

Tokenization

Splitting the text into tokens

Tokens

Words, parts of words or punctuation

Embedding/vectors

Each token ID is converted into a vector in embedding space

Embedding space

A multidimensional map that lets the model capture various patterns and relationships, grouping words based on patterns like words that often pair together

Transformer and attention

Model identifies which tokens matter most and maintains context across the sequence

Interpolation

When we estimate within our range of known data

Extrapolation

When we estimate outside our range of known data

Generator (GANs)

Receives random data vectors with certain statistical properties, uses the feedback from the discriminator to adjust and improve on its product

Discriminator (GANs)

Receives specified data vectors so it can learn patterns

Conditional GANs

GAN with an additional piece of information that we code into the model

Controllable GANs

Extension of a conditional GAN where we manage many features simultaneously